def main():
children = []
for cid, dev in enumerate(cuda.list_devices()):
t = threading.Thread(target=device_controller, args=(cid, ))
t.start()
children.append(t)
for t in children:
t.join()
print('ending gracefully')
def main():
children = []
for cid, dev in enumerate(cuda.list_devices()):
t = threading.Thread(target=device_controller, args=(cid,))
t.start()
children.append(t)
for t in children:
t.join()
print( 'ending gracefully' )
def __init__(self, gpuID=None, stream=None):
if gpuID is not None:
if gpuID < len(cuda.list_devices()) and gpuID >= 0:
cuda.close()
cuda.select_device(gpuID)
else:
raise ValueError('GPU ID not found')
if stream is None:
self.stream = cuda.stream()
else:
assert isinstance(stream, numba.cuda.cudadrv.driver.Stream)
self.stream = stream
self.blas = numbapro.cudalib.cublas.Blas(stream=self.stream)
self.blockdim = 32
self.blockdim2 = (32, 32)
import numpy as np
from math import ceil
import threading
from numbapro import cuda
print('System has %d CUDA devices' % len(cuda.list_devices()))
signature = 'void(int32[:], int32[:])'
def kernel(dst, src):
'''A simple kernel that adds 1 to every item
'''
i = cuda.grid(1)
if i >= dst.shape[0]:
return
dst[i] = src[i] + 1
# Numba compiler is not threadsafe
compiler_lock = threading.Lock()
def device_controller(cid):
cuda.select_device(cid) # bind device to thread
device = cuda.get_current_device() # get current device
# print some information about the CUDA card
prefix = '[%s]' % device
print(prefix, 'device_controller', cid, '| CC', device.COMPUTE_CAPABILITY)
import numpy as np
from math import ceil
import threading
from numbapro import cuda
print('System has %d CUDA devices' % len(cuda.list_devices()))
signature = 'void(int32[:], int32[:])'
def kernel(dst, src):
'''A simple kernel that adds 1 to every item
'''
i = cuda.grid(1)
if i >= dst.shape[0]:
return
dst[i] = src[i] + 1
# Numba compiler is not threadsafe
compiler_lock = threading.Lock()
def device_controller(cid):
cuda.select_device(cid) # bind device to thread
device = cuda.get_current_device() # get current device
# print some information about the CUDA card
prefix = '[%s]' % device
print( prefix, 'device_controller', cid, '| CC', device.COMPUTE_CAPABILITY )
max_thread = device.MAX_THREADS_PER_BLOCK
with compiler_lock: # lock the compiler